Automatic Differentiation

[pzivich]

This branch proposes an implementation of automatic differentiation to compute the bread, as mentioned in #23. While there are libraries, like JAX, that implement this, I do not want to add dependencies. Therefore, I have implemented a version of automatic differentiation.

• Develop autodiff functionality
• Develop tests for popular functions we want the derivative of
• Develop tests for all estimating equations and compare bread computations

Note that this branch is still under active development. I have not fully tested my implementation of computing the bread. Therefore, only use this branch for preliminary testing.

Note: I am planning on keeping the derivative approximation of the bread as the default. The exact derivatives are going to be an alternative option. Ultimately, any errors should lead the user to fall back to the approximation.

[pzivich]

The addition of ee_glm might make the AD implementation a little more difficult. Specifically, the AD will need to handle the polygamma and digamma functions and be able to get their derivatives

[pzivich]

Okay here is how I think I handle digamma and polygamma.

1. Write an internal polygamma function. This functions checks the input type. If it is a PrimalTangentPairs object, it calls the internal function on that object. Otherwise, it puts it into the scipy version of the polygamma function. Repeat for digamma.
2. Add internal polygamma function to PrimalTangentPairs that correctly creates each pair. Follow syntax of polygamma. Same for digamma.

This setup should handle the new ee_glm equations by correctly sorting out the polygamma pieces into the correct spots.

The downside to this is if users wanted to use the polygamma function, they would need to would need to use my internal version when pairing with AD. Thankfully this is all avoided when using numerical approximation

[pzivich]

Okay I believe to have fixed my NumPy woes. Now I need to create autodiff tests for all the available estimating equations (to be safe everything works with what I provide to users).